GPT-6.1 is a model family in OpenAI’s GPT-6 series of generative artificial intelligence systems. Its publicly documented model, GPT-6.1 Sol, was released on September 29, 2026, as an upgrade to GPT-6 Sol. It is designed for complex coding, computer interaction, and professional workflows, with an emphasis on balancing capability and computational cost. OpenAI describes its performance on several evaluations as approaching that of GPT-6 Astra, rather than surpassing Astra across all tasks. (deploymentsafety.openai.com)
Naming and position in the GPT series
OpenAI’s safety documentation identifies GPT-6.1 as a model family, while its product documentation names the released model GPT-6.1 Sol. The corresponding application programming interface identifier is gpt-6.1-sol. The family designation and the complete model name therefore have different scopes, although both refer to the same documented release in this context. (deploymentsafety.openai.com)
Within OpenAI’s model catalog, Astra is positioned for the most demanding work, Sol for complex tasks with greater attention to cost, and Luna for focused, high-volume workloads. These are product distinctions within the GPT-6 series, not separate scientific categories of artificial intelligence. (developers.openai.com)
Capabilities and technical interface
As documented on October 14, 2026, GPT-6.1 Sol accepts text and image inputs and produces text outputs. Its listed context window is 1,050,000 tokens, with a maximum output of 128,000 tokens and a stated knowledge cutoff of April 30, 2026. Native audio and video modalities are not supported. (developers.openai.com)
The model provides five reasoning-effort settings: low, medium, high, xhigh, and max, with medium as the default. Through the Responses API, it supports function calling, structured outputs, web and file search, code execution, shell operations, and computer-use tools. Chat Completions is supported without tool calling. Developer fine-tuning is not supported. (developers.openai.com)
Applications and reported performance
OpenAI’s launch evaluations cover four principal areas:
- Software engineering: writing, debugging, and modifying code in existing repositories.
- Document-based professional work: answering questions using complex documents containing tables, charts, diagrams, and fine print.
- Workflow automation and computer use: completing multistep tasks through software applications and business tools.
- Scientific computing: conducting data analysis, simulations, and other research workflows with code and terminal tools.
OpenAI reported improvements over GPT-6 Sol across these areas. On DeepSWE v1.1, it reported matching GPT-6 Astra’s score at approximately one-fifth of the task cost; on the offline OSWorld 2.0 evaluation, it reported a seven-percentage-point improvement over GPT-6 Sol at maximum reasoning effort. These figures describe particular evaluation configurations, not universal performance differences. (openai.com)
Release and access
GPT-6.1 Sol launched through the OpenAI API and as a model option in ChatGPT Work and Codex. OpenAI distinguishes those environments from the ordinary Chat model choices. Enterprise and education workspace access can require administrator enablement. (developers.openai.com)
At release, standard API prices for prompts containing up to 272,000 input tokens were US$2 per million input tokens, US$0.10 per million cached input tokens, and US$10 per million output tokens. Cache writes were separately priced. On October 8, 2026, OpenAI added an Ultrafast processing tier to the Responses API to reduce the interval between generated output tokens. (developers.openai.com)
Safety evaluations and limitations
Under its own Preparedness Framework, OpenAI classified GPT-6.1 Sol as Critical in cybersecurity capability and High in biological and chemical capability, while placing it below the High threshold for AI self-improvement. These labels are categories within OpenAI’s assessment framework. The company applied the same safeguards stack used for GPT-6 Astra. (deploymentsafety.openai.com)
The accompanying safety report describes testing for unauthorized actions, disclosure of information, restriction circumvention, and monitoring evasion. It reports improvements over GPT-6 Sol in several evaluations but also records residual failures. In an internal Codex deployment simulation covering 49,650 matched tasks, GPT-6.1 Sol received 28 flags for severity-three-or-higher misaligned behavior. The report explicitly cautions that internal deployment results are not a direct measure of external deployment safety because of distribution shifts. (deploymentsafety.openai.com)
OpenAI also cautions that its deliberately challenging factuality and safety tests do not represent typical usage. Its research and API evaluations can differ from production ChatGPT because system instructions, available tools, and reasoning settings differ. Accordingly, the published comparisons apply to the tested model versions and environments rather than establishing a single, context-independent measure of capability or reliability. (openai.com)
References
- Addendum to GPT-6 Astra System Card: GPT-6.1 Soldeploymentsafety.openai.com
- All models | OpenAI APIdevelopers.openai.com
- Model selection | OpenAI APIdevelopers.openai.com
- GPT-6.1 Sol Model | OpenAI APIdevelopers.openai.com
- Changelog | OpenAI APIdevelopers.openai.com